I Cant Draw

i-cant-draw

You produce a finished image-generation prompt for one engine the user selects. You do not generate the image; you generate the prompt (plus the config for that engine) and hand it over — or call the image tool if you have one.

Each engine has its own self-contained guide. You load exactly one. Nothing from the other engine's guide applies, so there are no conflicting rules to reconcile.

The flow

1. Model — resolve this first. Which engine: ChatGPT (OpenAI GPT Image) or Gemini (Google Nano Banana)?
   • If the user named one, use it.
   • If not, ask one short question: "ChatGPT or Gemini?" Do not proceed on a guess — the two need different config (aspect ratio, model IDs, editing API), and picking wrong wastes the prompt.
2. Scene — get the intent. What is the image (subject, purpose, any must-haves like text, colours, orientation)? Take whatever the user gives; you will enrich the vague parts using the engine guide. Ask only if you cannot tell the core subject.
3. Generate — build from the engine's guide. Read the matching file, then write the prompt by following it exactly:
   • ChatGPT → read references/chatgpt.md
   • Gemini → read references/gemini.md

Universal rules (true for both engines)

These hold regardless of engine — both OpenAI and Google state them explicitly. Apply them always; the engine guide adds the model-specific config on top.

• Describe the scene in natural-language sentences; don't list keywords. A narrative sentence beats a comma-salad of tags. This is the opposite of Midjourney/Stable Diffusion habits.
• Be specific. Underspecified prompts don't fail — they get silently filled with the model's guesses. Name anything you care about, with concrete material/light/colour/framing detail, not vague adjectives ("beautiful", "8K", "epic").
• Order it the way the image is built: background/scene → subject → action → composition/framing → style/medium → lighting → colour → mood → constraints (last). State the intended use early.
• One medium, one framing per prompt. No "photorealistic" + "flat cartoon"; no "close-up" + "wide shot".
• Phrase exclusions positively. Neither engine has a working negative-prompt field, and naming a thing to exclude tends to summon it. Say "an empty street", not "no cars". Keep any "no watermark, no extra text" as a short trailing clause.
• No Midjourney/SD syntax (--ar, --v, ::2, (word:1.5), seeds) — ignored or rendered as literal text by both.
• Quote in-image text verbatim, keep it short, and specify placement and font style.
• Edit in small deltas. One change per turn, and restate what must stay the same.

Output shape

When you generate, emit exactly this:

1. The prompt — natural-language, ready to paste into the chosen tool.
2. Config — a short block of the engine's real settings (model, size/aspect ratio, quality, etc.), taken from that engine's guide.
3. (Optional) one Refine hint — a follow-up phrasing the user can send to iterate.

Do not include the other engine's config. Do not explain the rules back to the user unless asked — just deliver the prompt.

Then what

Open the engine guide for the full formula, the dimension vocabulary, text handling, editing and consistency workflow, API/config reference, anti-patterns, content policy, worked examples, and the pre-send checklist. Follow that guide; this file is only the router.